Improving Time-based Predictions in Predictive Process Monitoring

Predictive process monitoring (PPM) aims to anticipate the future behavior of ongoing process instances, with time-related predictions such as remaining time or next-event timestamps being particularly valuable for proactive decision-making. Despite the success of deep learning approaches, prediction errors remain substantial, limiting their practical applicability. While prior research has extensively examined bucketing, trace encoding, and model design, the role of feature engineering remains largely underexplored. This PhD research addresses this gap by systematically investigating how process-specific knowledge can be embedded into engineered features and how this impacts predictive accuracy for time-related targets. It draws upon the methodology of algorithm engineering to develop and empirically evaluate knowledge-informed feature designs. Complementary aspects of process knowledge, such as seasonality, working-time constraints, and context, are operationalized into feature representations. These mechanisms are brought together in a synthesis study to derive empirically grounded design principles on how process knowledge improves time prediction in PPM.